def test_quality(t_true,
y_true,
pred,
time_grid=np.linspace(0, 300, 30, dtype=np.int),
concordance_at_t=None,
plot=False):
# get survival proba for time_grid
all_surv_time = pd.DataFrame()
for t in time_grid:
surv_prob = np.exp(-1 * np.power(t / (pred[:, 0] + 1e-6), pred[:, 1]))
all_surv_time = pd.concat([all_surv_time, pd.DataFrame(surv_prob).T])
all_surv_time.index = time_grid
ev = EvalSurv(surv=all_surv_time,
durations=t_true,
events=y_true,
censor_surv='km')
dt_c_index = ev.concordance_td('antolini')
int_brier_score = ev.integrated_brier_score(time_grid)
int_nbill = ev.integrated_nbll(time_grid)
if plot:
fig, ax = plt.subplots(1, 3, figsize=(20, 7))
d = all_surv_time.sample(5, axis=1).loc[1:]
obs = d.columns
for o in obs:
ax[0].plot(d.index, d[o])
ax[0].set_xlabel('Time')
ax[0].set_title("Sample survival curves")
nb = ev.nbll(time_grid)
ax[1].plot(time_grid, nb)
ax[1].set_title('NBLL')
ax[1].set_xlabel('Time')
br = ev.brier_score(time_grid)
ax[2].plot(time_grid, br)
ax[2].set_title('Brier score')
ax[2].set_xlabel('Time')
plt.show()
if concordance_at_t is not None:
harell_c_index = concordance_index(
predicted_scores=all_surv_time.loc[concordance_at_t, :].values,
event_times=t_true,
event_observed=y_true)
return pd.DataFrame([{
'harell_c_index': harell_c_index,
'dt_c_index': dt_c_index,
'int_brier_score': int_brier_score,
'int_nbill': int_nbill
}])
else:
return pd.DataFrame([{
'dt_c_index': dt_c_index,
'int_brier_score': int_brier_score,
'int_nbill': int_nbill
}])
def get_metrics(val_data, surv_pred, time_grid):
ev = EvalSurv(surv_pred, val_data['t'], val_data['y'], censor_surv='km')
return pd.DataFrame([{
'dt_c_index':
ev.concordance_td('antolini'),
'int_brier_score':
ev.integrated_brier_score(time_grid),
'int_nbill':
ev.integrated_nbll(time_grid)
}])
def Bootstrap(self, surv, event: list, duration: list):
np.random.seed(42) # control reproducibility
cindex, brier, nbll = [], [], []
for _ in range(self.bootstrap_n):
sampled_index = choices(range(surv.shape[1]), k=surv.shape[1])
sampled_surv = surv.iloc[:, sampled_index]
sampled_event = [event[i] for i in sampled_index]
sampled_duration = [duration[i] for i in sampled_index]
ev = EvalSurv(sampled_surv, np.array(sampled_duration),
np.array(sampled_event).astype(int), censor_surv='km')
time_grid = np.linspace(min(sampled_duration), max(sampled_duration), 100)
cindex.append(ev.concordance_td('antolini'))
brier.append(ev.integrated_brier_score(time_grid))
nbll.append(ev.integrated_nbll(time_grid))
return cindex, brier, nbll
# Training ======================================================================
epochs = args.epochs
callbacks = [tt.callbacks.EarlyStopping()]
verbose = True
log = model.fit(x_train,
y_train,
batch_size,
epochs,
callbacks,
verbose,
val_data=val,
val_batch_size=batch_size)
# log = model.fit(x_train, y_train_transformed, batch_size, epochs, callbacks, verbose, val_data = val_transformed, val_batch_size = batch_size)
# Evaluation ===================================================================
_ = model.compute_baseline_hazards()
surv = model.predict_surv_df(x_test)
ev = EvalSurv(surv, durations_test, events_test, censor_surv='km')
# ctd = ev.concordance_td()
ctd = ev.concordance_td()
time_grid = np.linspace(durations_test.min(), durations_test.max(), 100)
ibs = ev.integrated_brier_score(time_grid)
nbll = ev.integrated_nbll(time_grid)
val_loss = min(log.monitors['val_'].scores['loss']['score'])
wandb.log({'val_loss': val_loss, 'ctd': ctd, 'ibs': ibs, 'nbll': nbll})
wandb.finish()
callbacks = [tt.callbacks.EarlyStopping()]
log = model.fit(x_train, y_train, batch_size, epochs, callbacks, val_data=val)
surv = model.predict_surv_df(x_test)
ev = EvalSurv(surv, durations_test, events_test, censor_surv='km')
result = pd.DataFrame([[0]*8],columns=["random","model","hiddens",
"lr","scalers","c-index","brier","nll"])
result["c-index"] = ev.concordance_td('antolini')
print(ev.concordance_td('antolini') )
time_grid = np.linspace(durations_test.min(), durations_test.max(), 100)
result["brier"] = ev.integrated_brier_score(time_grid)
print(ev.integrated_brier_score(time_grid) )
result["nll"] = ev.integrated_nbll(time_grid)
result["lr"] = lr
result["model"] = mod
result["scaler"] = scale
result["random"] = seed
result["hiddens"] = dim
results = pd.concat([results,result],ignore_index=True)
results.to_csv(os.path.join(outpath,"results"))
pc_col = ['PC'+str(i) for i in range(x_train.shape[1])]
pc_train = pd.DataFrame(x_train,columns = pc_col)
plot_survival(pc_train,pc_col,model,outpath,duration=100)